SkyPilot raises $20M to unify fragmented AI compute across clouds and optimize GPU usage

SkyPilot raises $20M to be the Switzerland of AI compute

"If you spend $100 million a year on GPUs, we help squeeze out more than 10% more utilisation," SkyPilot says. That is a $10 million saving from efficiency alone. The startup just raised $20 million to become the Switzerland of AI compute. Its backers include the chief executives of Databricks, Vercel, Replit, and Hugging Face. The company came out of stealth on Tuesday. The founders built it at UC Berkeley, in the same lab that produced Spark, Databricks, and Anyscale. The team includes Zongheng Yang and Databricks co-founders Ion Stoica and Scott Shenker. The star-studded cheque Lux Capital led the $20 million seed, Fortune first reported. Coatue, Amplify, Foundation, Race, and The House Fund also wrote checks. The angel list is the tell. It runs from Databricks' Ali Ghodsi and Google's Jeff Dean to the bosses of Vercel, Replit, and Hugging Face. They are betting on a layer Nvidia already wanted. It bought Run:ai for about $700 million in 2024 to solve a version of this problem. Analysts expect the AI orchestration market to grow from about $14 billion this year to more than $60 billion by 2034. One pool of scattered compute The problem SkyPilot targets is fragmentation. GPU demand far outstrips supply. So every AI team now calls five or ten providers on day one just to scrape together chips, as SiliconAngle explains. Then it struggles to use them together. SkyPilot is a control plane that pools it all into one interface, across 20-plus clouds, neoclouds, Kubernetes, and Slurm. It picks the most available hardware, packs workloads onto idle chips, and moves jobs without a rewrite. The open-source version has passed 14 million downloads. Top deployments already run over 10,000 GPUs. Why neutrality is the moat SkyPilot answers to no single cloud or chipmaker. It counts Nebius and CoreWeave as partners, not rivals. Its edge is simple: it sends your workload wherever it runs cheapest. That reframes the question hanging over the industry: can AI firms make money? Yang points to Cursor, whose margins were negative until it stopped renting a rival's models and trained its own. He calls that shift "custom intelligence," now cheaper thanks to open-weight models that rank near GPT and Claude. The catch There is an obvious hole. The code has sat free on GitHub for years, so what stops a customer using it without paying? Lux's Brandon Reeves has an answer: the free version is, he says, "probably like 1% of the way done." The deeper bet is Stoica himself. His Berkeley lab already produced two multibillion-dollar companies. That makes it a magnet for the students who build the next one. SkyPilot is selling a promise as much as a product. Whoever controls the compute layer, not any one chip, may win the next phase of AI.

SkyPilot nabs $20M to ease AI infrastructure management

Infrastructure automation startup SkyPilot Inc. launched today with $20 million in seed funding. Lux Capital led the raise with participation from several other institutional backers. The funds were joined by Databricks Inc. Chief Executive Officer Ali Ghodsi, Google LLC chief scientist Jeff Dean and other prominent angel investors. San Francisco-based SkyPilot commercializes an open-source project of the same name. It was developed at UC Berkeley by the company's founding team, which includes Databricks co-founder Ion Stoica. The software makes it easier for enterprises to manage the infrastructure they use in artificial intelligence projects. Companies often mix and match different types of AI infrastructure. For example, an enterprise might run inference workloads in the cloud while keeping its training environments and the associated datasets on-premises. In many cases, each infrastructure environment contains different chips and management tools. Managing hardware assets with diverging technical parameters can be highly challenging. Usually, moving workloads between different infrastructure types requires extensive code changes. Additionally, automation workflows that run reliably in one environment might not work well in another. SkyPilot's eponymous open-source tool eases the task. It enables developers to manage infrastructure environments with different configurations through a single interface. According to the company, the software also automates several common maintenance tasks. Developers can sometimes make an AI cluster more efficient by changing which server runs what workload. If a machine with 10 gigabytes of memory runs a 6-gigabyte workload, an engineer could replace it with two 5-gigabyte workloads. That optimization approach is known as bin packing. SkyPilot says that its software performs the task automatically to save time for users. When an AI workload's hardware requirements increase, the software automatically provisions additional infrastructure. It can also fix certain types of technical issues without human input. SkyPilot monetizes the software with a paid version called the SkyPilot Platform. It includes a tool for creating inference sandboxes, virtual machines in which AI agents can run code without creating cybersecurity risks. Developers may customize the virtual machines by adding in components such as programming tools. Launching an inference sandbox can take a significant amount of time in some cases. According to SkyPilot, its platform speeds up the task by loading virtual machines before they're needed. That enables the software to activate a new sandbox in under a second when it's requested by an AI agent. The SkyPilot Platform also speeds up AI training runs. The company says that it can launch graphics card clusters with 5,000 chips in under a minute. For added measure, a tool called the SkyPilot GPU Manager regularly checks the accelerators' health and fixes any technical issues it finds. "Every organization is building custom intelligence around its own data and domains," said SkyPilot co-founder and CEO Zongheng Yang. "The challenge is that the AI compute needed to build it is fragmented across clouds. SkyPilot gives frontier AI teams a single platform to manage that infrastructure so they can build custom intelligence faster." SkyPilot will use its funding round to enhance the paid and open-source versions of its platform. The company also plans to grow its headcount.